Autonomous Supply Chain Control Tower
An AI-native supply chain control tower: a fully autonomous 13-agent LangGraph system that turns any dropped CSV into pipelines, ML features, a retrained model, and a refreshed dashboard.
Disruptions resolved before you wake up.
Most supply-chain platforms need humans to write ETL, tune models, and watch pipelines. This one does all of it autonomously: drop a CSV into a watched folder and the system infers the schema, builds the pipeline, engineers ML features, retrains the delay-prediction model, promotes it to production, and updates Grafana, with zero human intervention.
A 13-agent control plane over a medallion lakehouse.
The platform is built in five layers, data enters through three sources, flows through a medallion lakehouse into the ML models, and is served to a dashboard and API. A control plane of 13 deepagents (LangGraph) agents observes every layer and issues corrections.
A DataIngestionAgent uses an iterative loop to infer schema and validate the loader (up to 3 auto-fix attempts); MLflowGuardian detects drift and promotes better models via a Claude tool-use decision with hard floors (`roc_auc >= 0.60`); DagsterGuardian re-runs failed pipelines and escalates after repeated failures.
Structured outputs, not regex JSON parsing.
When a cross-agent failure is detected, an Orchestrator LangGraph graph with three specialist sub-agents diagnoses root cause and issues structured `OrchestrationResult` corrections, no brittle regex JSON parsing. New features proposed by Claude pass an 8-gate validation sandbox (distribution sanity, variance, novelty) before being merged into the next training run.
Live, tested, and continuously self-healing.
A DashboardAgent pushes new Grafana panels every ten minutes from structured Claude output. The system ships with a live demo, 356 passing tests, and all 13 agents reporting healthy.